<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Vinit Tople on Medium]]></title>
        <description><![CDATA[Stories by Vinit Tople on Medium]]></description>
        <link>https://medium.com/@vinittople?source=rss-3c600bf0ea6b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*pIQ7vxE46TYsQ7Mi</url>
            <title>Stories by Vinit Tople on Medium</title>
            <link>https://medium.com/@vinittople?source=rss-3c600bf0ea6b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 08:07:08 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@vinittople/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[We’re In the Same AI Theater But Watching Different Movies: 5 Points of Conflation]]></title>
            <link>https://medium.com/write-a-catalyst/were-in-the-same-ai-theater-but-watching-different-movies-5-points-of-conflation-0dfe73079bc4?source=rss-3c600bf0ea6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/0dfe73079bc4</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Vinit Tople]]></dc:creator>
            <pubDate>Thu, 30 Apr 2026 04:04:53 GMT</pubDate>
            <atom:updated>2026-04-30T04:04:53.011Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c44GnWt3xfolMiFJCQyDJQ.png" /><figcaption>Generated by AI</figcaption></figure><p>Just got back from Google Next 2026. If there’s one thing the hallway conversations confirmed, it’s this: we’re all sitting in the same AI theater, but we’re watching very different movies.</p><p>AI has moved well past the hype era. The impact is real and increasingly non-linear. Productivity gains measured in multiples, not percentages. The question has shifted from “should we invest in AI?” to “how do we capture its full potential?” That means meaningful changes to platforms, processes, and how teams work together. But what changes exactly and in what sequence? Some answers exist. Some are genuinely unknown. What’s clear, though, is that <em>the bulk of what’s stalling enterprises isn’t a shortage of capability or solutions — it’s coordination and consensus</em>.</p><p>Across dozens of conversations at Next, the same pattern kept surfacing: enormous organizational energy is burning on churn — not because people disagree, but because they’re unknowingly having different conversations using the same vocabulary. And the root cause? The AI landscape is evolving fast enough that people are operating at wildly different altitudes: some deep in production, some mid-flight with pilots, some still planning the runway. And they’re often speaking entirely different languages.</p><p>While it’s impossible to address every layer of maturity in one article, here are five points of confusion I kept encountering. Not definitions — just five areas worth pausing on because recognizing that confusion is itself a step toward clarity.</p><p><strong>1. Agentic Coding vs. Agentic Applications</strong></p><p>These get conflated constantly, and clarity here matters. Agentic coding is using AI to write software — whether that software is traditional or agentic. Agentic applications are the systems where LLMs can plan, reason, call tools, self-reflect, and act autonomously. Both carry real implications for platforms, processes, and people. But they’re different challenges with different risk profiles, and treating them as one conversation guarantees a muddled one.</p><p><strong>2. Context Management vs. Context Engineering</strong></p><p>Every executive has heard that “context is everything” in AI. That’s true but in two distinct ways and those two keep getting tangled. Context management is an enterprise-level concern: the infrastructure of connectivity through protocols like MCP, knowledge graphs, semantic data enrichment layers and more. Context engineering on the other hand operates at the individual agent level — it is the craft of making the right information available, in the right amount, in the right sequence, at the right time. One is infrastructure strategy. The other is design discipline. They’re related, but they’re not the same conversation. Furthermore, some refer to context engineering as agent engineering while others treat it differently.</p><p><strong>3. Desktop Agents vs. Production Agents</strong></p><p>The barrier to building agents (or any software) collapsed last year, and that’s genuinely exciting. Tools like Claude Code turned a wide swath of people into agent builders almost overnight. MCP gave those agents connectivity. Suddenly, you can spin up an agent on your laptop that does real work.</p><p>But there’s a meaningful gap between these “agents” which are deployed on a laptop versus the agents which are deployed in production. Production-grade agents require governance, security controls, observability, evaluation pipelines, and CI/CD tollgates — fundamentally different software profiles. Too often, these are discussed as though solving for one solves for the other. That’s a highly risky assumption.</p><p><strong>4. Prototyping Isn’t Failure — It’s the Starting Line</strong></p><p>The “95% of AI projects don’t get past prototype” statistic has made prototyping feel like a dirty word. In the world of agents, that framing misses the point. Traditional software is deterministic — you can lay out requirements upfront and fund your way to an outcome. Agentic systems are probabilistic. You cannot write detailed specs before you build, because you don’t yet know what you’ll need.</p><p>A strong prototype gets you to say 60% (just an example) of a solution. That’s not failure — that’s where the real work begins. The journey from 60% to your acceptable threshold is where you discover true requirements, often expressed not as specification documents, but as “agentic evaluations”. Agentic Evaluations define what probability is acceptable for each objective of what the agent does. Engineers shouldn’t treat that 60% as a victory lap. Business leaders shouldn’t call it a wasted investment. It’s a necessary milestone, and the sooner both sides internalize that, the faster things actually move.</p><p><strong>5. The Word “Agent” Itself</strong></p><p>LLMs. RAG based systems. Tool-calling LLMs. Planning engines. Multi-agent orchestrations. Self-improving loops — the industry calls all of them “agents”. No single agreed-upon definition exists. This alone accounts for a staggering amount of miscommunication. When someone says “we need an agent for X,” five people in the room picture five different architectures with five different risk profiles. Before debating the solution, it’s worth pausing to align on what you’re discussing.</p><p><strong>So What Do You Do With This?</strong></p><p>First, recognize the points of conflation with a sense of empathy and humility. Then seek to understand or to be understood. And the above list is only a start. There is no silver bullet to tackle this but inaction is also chaos and stalemate. Each person, starting with leaders, has to remediate. Towards that, I’ve come to think that in AI, everyone carries three roles — regardless of seniority.</p><ul><li><strong>Your day job:</strong> the actual work you’re accountable for delivering.</li><li><strong>Student (constant learning): </strong>because the landscape shifts weekly, and the need for continuous learning is materially higher than it’s ever been.</li><li><strong>Teacher (bringing others along):</strong> This is the role that is often understated. Unless every company becomes a one-person operation, you still have to bring people along. Alignment requires sharing what you’ve each learned and building clarity where confusion exists. Without that, you talk past each other indefinitely, <em>and the organization with less of that friction moves past yours. </em>Teaching can happen at an individual, org, or enterprise level.</li></ul><p>So how do you carve out time for the learning and teaching? No prizes for guessing — by using AI and agents to compress the day job. The tools are already here.</p><p>One last observation. Across nearly every session at Google Next, the leaders and organizations making real progress echoed the same point: <em>you can’t learn to swim by watching tutorials. </em>Get in the water. Prioritize transformative initiatives that create your moat, and figure things out along the way. There are enough rafts for those willing to jump in.</p><p><em>The views expressed in this article were mine alone and do not necessarily reflect the views of my current or former employers, or any organizations with which I’ve been affiliated.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0dfe73079bc4" width="1" height="1" alt=""><hr><p><a href="https://medium.com/write-a-catalyst/were-in-the-same-ai-theater-but-watching-different-movies-5-points-of-conflation-0dfe73079bc4">We’re In the Same AI Theater But Watching Different Movies: 5 Points of Conflation</a> was originally published in <a href="https://medium.com/write-a-catalyst">Write A Catalyst</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Agent Architect: The Future-Proof Role for the AI Era]]></title>
            <link>https://medium.com/aimonks/agent-architect-the-future-proof-role-for-the-ai-era-8915d45a1fcc?source=rss-3c600bf0ea6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/8915d45a1fcc</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Vinit Tople]]></dc:creator>
            <pubDate>Tue, 27 May 2025 14:17:25 GMT</pubDate>
            <atom:updated>2025-07-24T13:28:50.322Z</atom:updated>
            <content:encoded><![CDATA[<p>SDEs in 90s, PMs in 2000s, Data Scientists in 2010s. Now it’s Agent Architects.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kltp0KhyBPjgdUzASqKwtA.jpeg" /><figcaption>Generated by DALL-E</figcaption></figure><p>If you’re not a member, you can read this for free <a href="https://medium.com/@vinittople/8915d45a1fcc?source=friends_link&amp;sk=187560b4de1b9b99a8cc8d1a2df18efc">here</a>.</p><p><em>The views expressed in this article are mine alone and do not necessarily reflect the views of my current or former employers, or any organizations with which I’ve been affiliated.</em></p><p>Over the past two decades, we’ve mastered the art of building software at scale — refining playbooks for efficiency, reliability, and control. Even as AI entered the enterprise, we adapted, embedding machine learning into workflows with clear inputs, deterministic triggers, and measurable outputs. But agentic systems break these assumptions. These aren’t static pipelines — they’re dynamic entities: reasoning, evolving, collaborating, and acting autonomously. They don’t just scale code — they scale judgment. And the architectural shift this demands is non-trivial. As organizations rush to build and deploy agents, they’re confronting a sobering reality: existing roles and design paradigms fall dramatically short. What we now need is a new kind of architectural thinking — and a new kind of architect.</p><h3>So, What’s ‘Net New’ With Agentic Systems?</h3><h4>New Architectural Primitives</h4><p>Agentic systems introduce radically new constructs: autonomous agents, agent teams, complex graphs, mutable memory, dynamic tool-calling, and continuous evaluators. These primitives represent a fundamental shift from traditional software paradigms:</p><ul><li><strong>Agents:</strong> Stateful, autonomous, capable of recursive planning.</li><li><strong>Graphs:</strong> Complex agent interactions resembling human-like collaboration.</li><li><strong>Memory:</strong> Mutable state management, continually evolving.</li><li><strong>Evaluators:</strong> Real-time metrics guiding autonomous adjustments.</li></ul><h4>Advanced Planning &amp; Reasoning</h4><p>Agents don’t merely execute; they reason, reflect, self-correct, and recursively plan their actions. Deciding when to remain deterministic versus leveraging autonomy requires nuanced, high-judgment design decisions.</p><h4>Multi-Agent Orchestration</h4><p>It’s a common misconception that the Large Language Model (LLM) alone determines agentic outcomes. In reality, outcomes hinge significantly on how agents are orchestrated:</p><ul><li>Clearly defined roles, responsibilities, and boundaries.</li><li>Precise agent-to-agent communication protocols.</li><li>Effective goal and task delegation among agents.</li></ul><h4>Memory as an Architectural Primitive</h4><p>Unlike traditional software, agent memory isn’t static. Each interaction updates memory, subtly and continuously shifting agent behavior. Unlike static databases, mutable memory isn’t easily version-controlled. This dynamism can introduce unpredictable drifts — without adequate tooling, metrics, or governance.</p><h4>Governance by Design</h4><p>Governance isn’t an afterthought — it’s foundational. Early-stage governance decisions — toll gates, guardrails, observability, and explainability — must be integral to initial designs. Reactive governance can’t salvage fundamental design oversights.</p><h3>Why Traditional Roles Fall Short</h3><p>Each traditional role — Software Architect, Data Scientist, Product Manager — excels individually yet struggles profoundly when tasked with designing agentic systems:</p><ul><li><strong>Software Architects:</strong> Skilled in deterministic, scalable systems; inexperienced in probabilistic, recursive reasoning architectures.</li><li><strong>Data Scientists:</strong> Experts in model tuning; unfamiliar with multi-agent orchestration and nuanced memory management.</li><li><strong>Product Managers:</strong> Strong in business alignment; typically lack deep technical insight into agent interactions and governance trade-offs.</li></ul><p>None individually — or even collectively — possess the hybrid skill set necessary to design, scale, and govern complex agentic systems at enterprise scale.</p><h3>The Rise of the Agent Architect</h3><p>Here emerges the Agent Architect — a cross-disciplinary role purpose-built for this agentic paradigm shift. The Agent Architect bridges the technical, product, and governance realms, uniquely positioned to navigate agentic complexity:</p><ul><li>Designs robust agent graphs and effective orchestration strategies.</li><li>Establishes memory strategies to balance mutability with stability.</li><li>Embeds governance deeply into architecture rather than reacting post-deployment.</li></ul><p>It’s inevitable that academia will eventually formalize this role, but for now, forward-looking enterprises must proactively nurture this talent internally.</p><h3>Misconceptions and Shortcuts (What to Avoid)</h3><p>Before we continue, let’s confront a common pushback: “Isn’t there already abundant training?” Yes, but training today splits awkwardly into two extremes — high-level strategic overviews for executives, or granular, introductory coding tutorials for developers. Critically missing is a middle ground: training at the architecture level, specifically tailored for product managers and aspiring Agent Architects.</p><p>Moreover, learning to build one agent is vastly different from architecting an entire fleet of governed, scalable agents. Deploying hundreds of agents in regulated production environments demands sophisticated orchestration and meticulous governance — something entirely underrepresented in current training resources.</p><h3>Architecting for Scale: Navigating the Complexity Chasm</h3><p>We’re facing what I term the “Agent Complexity Chasm”: the enormous gap between a functional prototype and a governed, production-ready agent ecosystem. Without dedicated Agent Architects, organizations risk chaotic agent sprawl, costly governance mistakes, or catastrophic failures — potentially eroding years of progress and public trust.</p><p>As enterprises pivot toward agentic systems, another reality becomes clear: traditional roles must adapt significantly. Take Product Managers: while this role has flourished over the past two decades — even for those minimally technical — that approach simply won’t work in AI-driven products. AI product management demands deep architectural understanding and technical fluency. Enterprises and Product Managers alike must acknowledge this shift early, or risk costly recalibrations later.</p><h4>Enterprises: What Are The Next Steps</h4><p>Recognizing and formalizing the Agent Architect role isn’t merely about a new title; it’s a strategic imperative:</p><ul><li>Clearly define the role’s responsibilities — publish these widely.</li><li>Pilot small-scale projects quickly to validate architectural patterns.</li><li>Codify successful pilots into reusable blueprints (graphs, memory patterns, guardrails).</li><li>Leverage the industry learnings. Learning from your own experience (mistakes) is one option but keep an eye out for the right developments. For example, currently, the only companies building agents at scale are those with Agent Builder product offerings. They’re the only ones with a vantage point that caters to this need.</li></ul><h4>Employees: How to Thrive</h4><p>Avoid chasing every flashy new framework. Instead, focus deeply on system-level skills:</p><ul><li>Master interactions between agents, memory management, evaluators, and governance guardrails.</li><li>Approach governance as code — build reusable, standardized guardrail modules.</li><li>Document real-world deployment lessons — especially failures — to build credibility.</li><li>Cultivate expertise in observability, evaluation, and systemic reasoning rather than mere coding.</li></ul><h3>Time for Action</h3><p>The true measure of sophistication is no longer a single clever agent; it’s the ability to orchestrate, scale, and govern entire fleets of autonomous agents seamlessly. Enterprises and practitioners who master this challenge early will dominate the next generation of AI-driven business.</p><p>The Agent Architect sits precisely at this crucial intersection — bridging experimental prototypes with the complex realities of large-scale deployment. By embracing this role now, we not only future-proof our organizations but position ourselves at the forefront of the next transformative wave in technology.</p><p>Are you already encountering the Agent Complexity Chasm? I’d love to hear your thoughts and experiences. My LinkedIn is <a href="https://www.linkedin.com/in/vinit-tople-9870a43/"><em>LinkedIn</em></a><em>. Also, Feel free to join my LinkedIn </em><a href="https://www.linkedin.com/groups/14580877/"><em>Group</em></a><em> to get articles on AI.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8915d45a1fcc" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aimonks/agent-architect-the-future-proof-role-for-the-ai-era-8915d45a1fcc">Agent Architect: The Future-Proof Role for the AI Era</a> was originally published in <a href="https://medium.com/aimonks">𝐀𝐈 𝐦𝐨𝐧𝐤𝐬.𝐢𝐨</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Yelps and The TripAdvisors: How Their Moat Is Fractured With AI Agents]]></title>
            <link>https://medium.com/write-a-catalyst/how-ai-agents-are-fracturing-moats-and-reshaping-business-models-32dc9cf10d9e?source=rss-3c600bf0ea6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/32dc9cf10d9e</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Vinit Tople]]></dc:creator>
            <pubDate>Mon, 27 Jan 2025 16:53:03 GMT</pubDate>
            <atom:updated>2025-02-03T16:17:58.381Z</atom:updated>
            <content:encoded><![CDATA[<p>Access this article for free <a href="https://medium.com/@vinittople/32dc9cf10d9e?source=friends_link&amp;sk=72e894904ee67e214a81135c1b25c210">here</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QbLHj-t-7fFzpdIXtIAcRw.png" /><figcaption>Generated by DALL-E</figcaption></figure><p>If 2023 was the year of AI models, 2024 the year of AI infrastructure, then 2025 is undoubtedly the year of AI Agents — poised to transform businesses and their underlying business models. AI Agents will play the role of both employees and customers. As Microsoft AI CEO said — We’re building a new digital species and as NVDIA CEO said — now IT departments will be the new HR department, of the AI Agents. In this article, we explore the disruptive potential of AI Agents, as the new customer, on one specific business model: one-stop shops for customer reviews, particularly those that combine review aggregation with marketplaces for products or services.</p><h4><strong>A 25-Year-Old Success Story Under Threat</strong></h4><p>Over the past 25 years, platforms like Yelp, TripAdvisor, Glassdoor, Goodreads, TrustPilot, ZocDoc, Angie’s List, Groupon, Rotten Tomatoes, Fandango, and OpenTable have dominated their respective domains. These platforms became immensely successful by offering a seamless combination of customer reviews and related bookings or purchases. They are the quintessential one-stop shops in their respective domains.</p><h4><strong>What Built Their Moats?</strong></h4><p>For decades, these companies have fortified their leadership positions with robust barriers to entry — or “moats.” Two critical moats (among others) stand out:</p><p><strong>1. Volume of Customer Reviews</strong><br> Reviews have become the most influential factor in purchase decisions, surpassing even price, free shipping, and recommendations from family and friends. A high volume of reviews on these platforms creates trust, credibility, and a reliable experience for users.</p><p><strong>The Disruption:</strong> Agents can now scan reviews across multiple platforms and consolidate them into a single, cohesive summary. For example, instead of a user relying solely on Yelp to decide where to eat, an Agent could combine reviews from smaller, niche platforms like EatLocal, RestaurantFinder, and foodie blogs to deliver a recommendation such as, “The best-rated vegan café near you is GreenSprout Café, with 4.8 stars across 12 review sources.” This levels the playing field, giving smaller platforms and businesses the opportunity to compete by being grouped into the same consolidated view as the industry giants. Suddenly, the value of sheer review volume on a single platform becomes much less critical.</p><p><strong>2. Familiarity with User Experience (UX)</strong><br> Platforms like Yelp and TripAdvisor don’t necessarily offer the “best” UX, but users, over time, have gotten familiar with their interfaces. This familiarity reduces cognitive load and creates a level of loyalty that’s hard for new entrants to disrupt.</p><p><strong>The Disruption:</strong> AI Agents can provide a hyper-personalized, conversational, and dynamically adaptive experience versus (and in addition to) the traditional website UX which requires a user to figure out the navigation. In short, Agents can dramatically reduce the cognitive load spent on figuring out the navigation to the exact insight/info that the user is seeking. The cognitive then shifts to the insight/information evaluation instead. Imagine this scenario: instead of opening TripAdvisor, searching for parasailing options, and sifting through reviews and prices, a user could simply say, “Find me the best-reviewed parasailing trip in Oahu under $200 with availability for my dates.” The Agent would instantly scan multiple sources — including TripAdvisor, Groupon, and smaller providers — and return a curated, consolidated option that fits the user’s criteria. The entire experience is tailored and frictionless, bypassing the need to navigate through a specific platform’s interface. The Agent doesn’t just match what the website can do — it does more, and it does it faster. Simply put, as more users prefer the Agent rather than the platform, it renders the platform’s UX less relevant. Another critical moat is weakened.</p><h4><strong>What Does It Mean for The Space?</strong></h4><p>This space is inevitably set to be rocked and that will be manifested in a few different ways. Below, let’s look at how some of these might play out.</p><ol><li><strong>Leveling of the Playing Field</strong><br> With traditional moats like review volume and UX familiarity eroding, new entrants have a unique opportunity to compete on tangible factors like price, quality, and innovation. The market becomes less about who has been there the longest or what users are used to and more about who can provide the best tangible value — price and quality.</li><li><strong>Revenue Streams at Risk</strong><br> Many platforms rely heavily on traffic-driven revenue streams, such as cross-selling and upselling. If Agents direct users to specific services or products without showing additional options (e.g., bundling parasailing with snorkeling), these revenue streams could diminish significantly.</li><li><strong>Emergence of Aggregator Agents</strong><br> The role of “aggregation” may shift from websites to Agents. While some incumbents may evolve to become leading Aggregator Agents, this field is still wide open. For example, today Yelp’s app is the “go-to” resource for local business reviews. Tomorrow, it could be Yelp’s Agent which becomes the “go-to” resource. Success will depend on who can master this new paradigm first. And while incumbents have an edge, there is now an “opening” for others.</li><li><strong>Rise of B2B Partnerships with Agents</strong><br> Just as businesses integrate with marketplaces like Amazon or app ecosystems like iOS and Android, we will likely see an explosion of B2B partnerships between companies and Agents. These partnerships will create new ecosystems where sellers and service providers hook into Agents as distribution channels. There will likely be an emergence of new “protocols” or “standards” for these integrations. For example, when an Agent visits a website, the way the website “speaks (interacts)” with the agent will be completely different from how it does with the customer. Today, such “agents” are like website crawlers and they’re just presented with the text content of the website but soon, there will be more formal and richer ways in which each website will expose its content — not just text but APIs/services/hooks to take actions with it.</li><li><strong>Unanswered Questions</strong><br> Much remains uncertain. Will Agent integrations be easier or harder for sellers compared to current platforms? Will users fully embrace the shift to Agent-led interfaces, or will platforms and Agents coexist? What role will incumbents play — will they dominate as Agents or lose relevance entirely? Will the “integrations” between websites (or product/service providers) and Agents be easier than the integrations that exist today between them and leading apps? Who will master the Agent UX for the respective domains? The verdict is not out on any of these but all of this being discussed and evaluated for sure.</li></ol><h4><strong>Key Takeaways</strong></h4><p><strong>Broader Phenomenon: Collapsing Application Stacks: </strong>This disruption is just one example of a larger trend, as Satya Nadella aptly pointed out in his December 2024 interview: “The application stack is collapsing.” The moats of many software applications — the UX and embedded business logic — are being fractured, if not entirely dismantled. Just addition of an Agent UX to the existing offering will not suffice. The business model itself needs to be revaluated and adapted.</p><p><strong>“Aggregation” as an Asset — Maybe Not:</strong> User-generated content, <em>which is public</em>, will continue to be valuable but merely ‘aggregating’ it in one place may cease to be a major differentiator. Scanning the entire web has become rather too easy and will only get easier.</p><p><strong>Solidification of the “Remaining” Moat:</strong> All is not lost for these incumbent or “software stack” owners — far from it. <strong>Proprietary data and APIs</strong> (task completion and 3P integrations) will remain critical assets<strong>.</strong> So will the brand name and trust that is built over the years. Businesses with unique customer insights, purchase histories, or deeply integrated APIs will still maintain significant competitive advantages. However, this will require intentional investment in recognizing and deploying these differentiating assets.</p><p><strong>Adaptability as a Winning Strategy: </strong>It’s rather obvious but given the rapidly evolving landscape of Agents, adaptability will be a crucial determinant of the eventual winners. This is the classic case of “building the airplane while flying it”. For incumbents and new entrants alike, the uncertainty and ambiguity is both the challenge and the opportunity.</p><blockquote>This is an industrial revolution. If you’re not thinking about AI today and how it’s going to change your business, you’re doing it wrong.</blockquote><blockquote>— Arthur Mensch, CEO, Mistral AI</blockquote><p><em>I’m Vinit Tople, an Ex-Amazon Product Leader (in AI &amp; eCommerce) and now an Industry Speaker, Writer &amp; Strategist on AI. Feel free to join my LinkedIn </em><a href="https://www.linkedin.com/groups/14580877/"><em>Group</em></a><em> to get articles on AI or just directly connect with me on </em><a href="https://www.linkedin.com/in/vinit-tople-9870a43/"><em>LinkedIn</em></a><em>.</em></p><p><em>If interested, below are additional articles by me on AI.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=32dc9cf10d9e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/write-a-catalyst/how-ai-agents-are-fracturing-moats-and-reshaping-business-models-32dc9cf10d9e">The Yelps and The TripAdvisors: How Their Moat Is Fractured With AI Agents</a> was originally published in <a href="https://medium.com/write-a-catalyst">Write A Catalyst</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[In The Adoption Of AI, What Is A Winning Hand?]]></title>
            <link>https://medium.com/aimonks/which-organizations-will-win-the-ai-game-differentiation-in-the-age-of-commoditization-c5c71f8d7e7f?source=rss-3c600bf0ea6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/c5c71f8d7e7f</guid>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Vinit Tople]]></dc:creator>
            <pubDate>Sun, 22 Dec 2024 04:51:53 GMT</pubDate>
            <atom:updated>2025-01-20T20:24:18.326Z</atom:updated>
            <content:encoded><![CDATA[<h3>In The Adoption of AI, What Is A Winning Hand?</h3><p>Know where your organization stands in this rapidly evolving game.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1r-cgPBw9SPs0XE075Q-EQ.png" /><figcaption>Source: Britannica.com</figcaption></figure><p><em>I’m Vinit Tople, an Ex-Amazon Product Leader (in AI &amp; eCommerce) and now an Industry Speaker, Writer &amp; Strategist on AI. Feel free to connect with me on </em><a href="https://www.linkedin.com/in/vinit-tople-9870a43/"><em>LinkedIn</em></a><em>.</em></p><h3><strong>Background</strong></h3><p>In the world of business, adoption of AI is rapidly underway but that is a survival strategy, not a winning one. For now, it may seem like a winning strategy but very quickly (2026–2028 onwards), it will become clear who truly embraced and internalized AI versus dabbled in it to look relevant, go with the flow or purely just survive. To win, you need differentiation even in your AI adoption strategy. The adoption strategy is getting commoditized. Yes, already.</p><p>Before surfacing the differentiators, let’s understand where the commoditization is first. It is in the form of the three aspects — 1) the hardware, 2) the AI models and 3) the middleware — the glue that binds various components of an AI based system to turn it into an “agent” which is basically what everyone is building (Agents are the new “App”). So, coming back to commoditization — everyone is using the same hardware (dominated by NVIDIA) and the same set of models from Big Tech (Open AI, Google, Microsoft, Amazon etc.). All these models are trained on same publicly available information so there’s not much differentiating about them. Lastly, the middleware (or orchestration tools) is largely from LangChain or Haystack or home grown but it’s not a big differentiator.</p><p>So, if all those aspects are commoditized, <strong><em>what is your differentiating card in this game</em></strong><em>? </em>Well, it is —<strong> <em>how deeply and strategically you leverage your proprietary assets </em></strong>to complement the commoditized parts (ones mentioned earlier) of your AI adoption strategy. That depth or extent or combination of “proprietary plus commoditized components” will determine whether you’re build a winning hand or merely playing the game. And what exactly are your proprietary assets? They are — your <strong>proprietary data</strong> and your <strong>proprietary APIs</strong>. Let’s explore a little about the “why” first and then discuss the three levels of “depth or extent” in utilizing proprietary data and APIs to create meaningful differentiation.</p><h3>Your Two Hidden Cards: Path to Differentiation</h3><h4>1. Proprietary Data</h4><p>The data you own is uniquely yours. Using that proprietary data to train an off-the-shelf AI model (e.g., from OpenAI or Google), which is already trained with public data will create a “layer of differentiation”. It’s worth noting only 5%-30% of enterprise data has been traditionally put to use by them. The rest of the data, though collected is never used for building insights and intelligence to inform decisions or actions. This ‘locked and trapped goodness’ is ready and able to give you the edge in this intense race. The following two aspects will drive more clarity of this ‘goodness’:</p><ul><li><strong>Structured to Unstructured Data:</strong><br>Older machine learning models thrived on structured data — tidy, formatted fields like SKUs, dates, names, order numbers or predefined categories. Today’s AI models, especially LLMs, can ingest <em>unstructured</em> data. This includes data like customer feedback, emails, chat logs, internal documents, or even support tickets — its ‘free text’ as opposed to ‘structured’ fields. This opens up entirely new avenues for extracting intelligence from previously untapped data sources. You can read <a href="https://medium.com/@lifeingeneral20/bezoss-legendary-empty-chair-now-has-a-butt-in-seat-eae47e2ad026">this</a> article, which gives you an even more specific and prime example of this value proposition.</li><li><strong>Higher Volume and Granularity:</strong><br>Modern AI models process a vastly greater number of data points with finer granularity. For instance, consider a car insurance company investigating accident claims. Traditionally, they might rely on basic details such as the car model, location, and the description of the incident. Now, AI can incorporate far richer and more granular data points: real-time weather data at the time of the accident, precise GPS tracking of the vehicle’s movements, telemetry data from the car’s sensors (e.g., speed, braking patterns), and even voice recordings of the driver reporting the incident.</li><li>Imagine the AI identifying that accidents occur more frequently in a specific area due to poorly marked intersections during rainy weather. Such granular insights can inform not just claim resolutions but also proactive customer recommendations, like issuing alerts for drivers when approaching high-risk zones during rain. This depth of insight, drawn from proprietary and real-time data, turns a reactive claims process into a predictive and preventive service.</li><li>Similarly, in retail, consider the personalization of product recommendations. Instead of suggesting products based on purchase history alone, AI can now analyze the customer’s exact clickstream data — what they hovered over, the sequence of their actions, and even the time spent contemplating a purchase — combined with real-time contextual information like current weather or ongoing promotions. This level of granularity delivers hyper-personalized experiences that are impossible to achieve with surface-level, publicly available data.</li></ul><h4>2. Proprietary APIs</h4><p>2023 was about text based Q&amp;A with AI systems like ChatGPT, which answers the question from its memory or its intelligence — its brain. However, 2024 has been about attaching “senses, arms and legs” to this brain. These senses, arms and legs are nothing but APIs which enabled models like ChatGPT to first pause, collect necessary info, then respond based on that <em>real-time information </em>(vs. just from its static memory) and then take actual physical action (e.g. turn off a Smart Light or send an email or change a setting in a car) versus just give you a text-based answer. Just like your data, these APIs are uniquely yours, specific to your products, processes, employees, suppliers and customers and hence — a major lever of differentiation.</p><p>So, with the “Why” out of the way, let’s move to the “How” and more specifically — the “How far”. How far are you willing to go to invest in your proprietary data and APIs, specifically with the purpose of embedding them in your AI powered capabilities. This “extent” of investment is what will determine the extent of your differentiation. Below is a simple “Basic — Medium — Advanced” structure that can help you calibrate that extent of investment</p><h4>1. Basic: Point Solutions for Targeted Use Cases</h4><p>At this level, companies use existing data and APIs to address isolated problems. These are often department-specific applications that demonstrate AI’s potential but lack cross-functional integration. They’re also mostly to provide “intelligence” (e.g., fraud detection model) as opposed to agentic solutions capable of making a set of API calls to collect context on the fly or take actual actions.</p><ul><li><strong>Example:</strong> A retailer builds a chatbot to recommend products based on customer purchase history and inventory data. While effective, it’s a single-use solution limited to the marketing function.</li></ul><p><strong>Insights:</strong> This stage is quick to implement, making it ideal for companies new to AI. However, it often results in “AI silos,” where solutions cannot scale or interact across the enterprise. These type of applications, though an advancement will soon fall into “too little, too late” category.</p><h4>2. Medium: Enterprise Enablement</h4><p>Organizations at this stage consolidate data from multiple sources and functions enabling AI models to serve broader business objectives. Most importantly, these organizations have recognized that the major value proposition of new-age AI is in two aspects:</p><p><strong>Cross Pollination: </strong>This is about <em>identification of insights/intelligence which would otherwise have never happened</em>. For example, a cosmetics company leverages R&amp;D insights about customer skin type reactions to specific formulations. While these insights were originally developed for improving product efficacy, they are fed into marketing algorithms to identify and target customers with personalized product recommendations, drastically improving ad conversion rates. The takeaway is that R&amp;D and marketing rarely collaborate directly, but AI bridges the gap by unlocking shared intelligence from centralized data. The list of such examples is long (excluded from this article only to limit the reading time).</p><p><strong>Task Completion:</strong> The high order bit, the needle mover, apart from the intelligence is task completion and that needs API integrations. The organizations develop new APIs to enhance functionality and cross-functional utility of AI.</p><p><strong>In summary, </strong>Medium-level maturity highlights AI’s potential as a cross-functional enabler and an action enabler. However, it requires significant investment in data integration, API development, and organizational alignment.</p><h4>3. Advanced: Transformational AI</h4><p>The pinnacle of AI maturity involves treating data and APIs as strategic assets. Some key traits of these companies are</p><p><strong>Data Strategy: </strong>Companies build centralized data lakes or meshes, implement robust governance, and most importance design with AI application in mind. They even proactively capture new types of data or from previously unprioritized sources, knowing their new found use. This data readiness will be instrumental in extracting the most value from their AI investments and at optimal speed.</p><p><strong>API Strategy: </strong>New APIs will be identified and invested in, keeping in mind potential use cases unlocked with agentic solutions. Existing APIs will be made ready for ingestion into agentic solutions with the right meta data or wrappers.</p><p><strong>Other Enablers: </strong>The most advanced application of AI goes well beyond the data and API aspect and its discussion is beyond the scope of this article. That said, Data and APIs are an integral part of the foundational layer.</p><p><strong>Example:</strong> An automotive company develops a “Digital Car Companion.” By combining unstructured feedback, IoT telemetry, and external APIs (e.g., traffic, weather), the agent proactively suggests maintenance, optimizes routes, and automates driver preferences. Over time, the system evolves, using insights to influence manufacturing or service protocols. Another example could be a large financial institution consolidating all its data from retail banking to investment banking to wealth management with an objective to centralize all of it and make it available for building of AI solutions — ML or LLM based or agentic.</p><p><strong>In summary, </strong>Advanced maturity transforms AI into a competitive moat and will likely transform industries. While resource-intensive, the payoff includes exponential scalability, efficiency, and entirely new business models.</p><h3>Closing Thoughts: Playing to Win</h3><p>In the world of AI, the “common cards” of models, hardware, and orchestration tools have become the table stakes. Your proprietary data and APIs are the hidden aces that determine whether you’re merely playing the game — or winning it.</p><p>In the race underway, know where you stand and plan/prepare accordingly:</p><ul><li><strong>Basic:</strong> Targeted applications for quick wins but with limited data and APIs.</li><li><strong>Medium:</strong> Scale AI enterprise-wide with consolidated data and <em>new</em> APIs.</li><li><strong>Advanced:</strong> Redefine the game with centralized, well-governed data and APIs designed and built for AI based agentic solutions at scale and speed.</li></ul><p>In summary, the real differentiator is not whether you embrace AI but how ambitiously and thoughtfully you do so. Like poker, it’s all about how you strategically you play your unique cards in tandem with the common cards on the table.</p><p>=============================================================</p><p><strong><em>I value all comments — agreeing or dissenting, here or directly to me (over </em></strong><a href="https://www.linkedin.com/in/vinit-tople-9870a43/"><strong><em>LinkedIn</em></strong></a><strong><em>)</em></strong>. — Vinit Tople</p><p>Also, here’s another related article: <a href="https://medium.com/aimonks/why-executives-should-make-a-note-of-ai-agents-40dbdc1a2977">Why Executives Must Take Note of AI Agents</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c5c71f8d7e7f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aimonks/which-organizations-will-win-the-ai-game-differentiation-in-the-age-of-commoditization-c5c71f8d7e7f">In The Adoption Of AI, What Is A Winning Hand?</a> was originally published in <a href="https://medium.com/aimonks">𝐀𝐈 𝐦𝐨𝐧𝐤𝐬.𝐢𝐨</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Product Managers  — Wake Up To This New Responsibility]]></title>
            <link>https://medium.com/aimonks/new-core-for-pms-building-ai-features-b0f849df5e36?source=rss-3c600bf0ea6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/b0f849df5e36</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[product-development]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Vinit Tople]]></dc:creator>
            <pubDate>Thu, 14 Nov 2024 02:54:31 GMT</pubDate>
            <atom:updated>2024-11-17T16:38:03.314Z</atom:updated>
            <content:encoded><![CDATA[<h3>AI Product Managers — Wake Up To This New Responsibility</h3><p>A key difference between PM-ing traditional software products vs. AI powered offerings.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F0KG4gV5uv5Ykkfg_md6hQ.jpeg" /><figcaption>Generated by DALL-E</figcaption></figure><p><em>I’m Vinit Tople, an Ex-Amazon Product Leader (in AI &amp; eCommerce) and now an Industry Speaker, Writer &amp; Strategist on AI. Feel free to connect with me on </em><a href="https://www.linkedin.com/in/vinit-tople-9870a43/"><em>LinkedIn</em></a><em>.</em></p><p><strong><em>Target Audience for this article</em></strong>: Product Managers who wish to build AI powered offerings.</p><p>In traditional product management, data played a supportive, almost secondary role. It served two main purposes:</p><ol><li><strong>Extracting Business Requirements:</strong> Data was analyzed to identify patterns or trends that would help define the product requirements. Only “valid” data points — high-quality and relevant — influenced these requirements. Any “invalid” or low-quality data was excluded or earmarked for cleanup, keeping it separate from the product’s core logic. The PM’s focus was on defining the requirements, which insulated the product from any potential data quality issues.</li><li><strong>Handling Data Transactions:</strong> Once the product was live, data flowed through it as user transactions, and if data quality issues arose, PMs could chalk it up to “garbage in, garbage out.” Poor data quality might impact the user’s output, but it didn’t reflect poorly on the product itself. The product’s functionality — its “plumbing” or underlying business rules — remained unaffected because these were predefined, independent of data quality.</li></ol><p>In both cases, data quality issues were often someone else’s responsibility, whether business users, internal customers, or analytics teams. However, this paradigm shifts dramatically in AI product management.</p><h4>Data as the New “Plumbing” for AI Products</h4><p>In AI, the very “plumbing” — the business rules and logic of the product — is driven by data. Unlike traditional products, where a PM might define and own every business rule, AI products rely on models that learn patterns, business rules, and logic directly from data. So, rather than merely flowing through the product, <strong>data now forms the product</strong>. As a result, the quality, relevance, and structure of data directly shape how the AI model learns and behaves.</p><p>This shift places product managers at the center of data quality and sourcing, making data a primary responsibility rather than a secondary consideration. <strong>PMs are now deeply accountable for the end product’s accuracy, fairness, and performance, all of which hinge on data quality.</strong></p><h4>The Complexity of Data Sourcing</h4><p>Sourcing quality data is no simple task. AI PMs must be ready to engage in detailed discussions and make strategic decisions about data quality, covering areas such as:</p><ul><li><strong>Availability and Volume:</strong> Is there data available, relevant for our training? If yes, is there enough of it?</li><li><strong>Format and Structure:</strong> Is it in the right format, labeled correctly, and structured to fit the model’s requirements?</li><li><strong>Bias and Integrity:</strong> Does the data fairly represent all necessary user groups? Is it biased, or has it been compromised in any way?</li><li><strong>Internal vs. External Sources:</strong> Can the data be sourced internally, or does it need to be acquired? If external, what are the costs, and does it come with necessary permissions?</li><li><strong>Artificial Data Creation:</strong> For some scenarios of limited data, should synthetic data be created to supplement training?</li></ul><p>These are not mere technical details; they’re strategic considerations with major implications for the product’s success and user experience. It’s why AI PMs need to develop a “muscle” for data sourcing and quality, becoming comfortable with this new, complex role. While data scientists play a significant (often primary) part in data management, the PM’s involvement is essential in defining data requirements, ensuring alignment with product goals, and making decisions that balance cost, quality, and compliance.</p><h4>Quote</h4><p>I once read in some article — “For AI PMs, data is no longer an input — it’s the entire blueprint.” This new responsibility requires AI product managers to think critically about data in ways traditional PMs never had to, ensuring that data truly represents the product’s purpose and user needs.</p><p>In short, if you’re managing an AI-powered product, <strong>data quality isn’t just a back-end detail; it’s the product itself.</strong> Embrace data conversations, build this new skill, and be prepared to engage deeply with data strategy.</p><p>===============================================================</p><p><em>I love connecting with my readers so feel free to reach out on </em><a href="https://www.linkedin.com/in/vinit-tople-9870a43/"><em>LinkedIn</em></a><em>. Also, </em><a href="https://medium.com/@lifeingeneral20/subscribe"><em>subscribe</em></a><em> or follow if you like reading about AI.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b0f849df5e36" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aimonks/new-core-for-pms-building-ai-features-b0f849df5e36">AI Product Managers  — Wake Up To This New Responsibility</a> was originally published in <a href="https://medium.com/aimonks">𝐀𝐈 𝐦𝐨𝐧𝐤𝐬.𝐢𝐨</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[This Should be the #1 Application of Gen AI in your Company]]></title>
            <link>https://medium.com/@vinittople/bezoss-legendary-empty-chair-now-has-a-butt-in-seat-eae47e2ad026?source=rss-3c600bf0ea6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/eae47e2ad026</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[ceo]]></category>
            <dc:creator><![CDATA[Vinit Tople]]></dc:creator>
            <pubDate>Mon, 30 Sep 2024 00:53:01 GMT</pubDate>
            <atom:updated>2025-01-04T18:17:35.921Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bs1LLO_ecSNoQXc6H6unVg.jpeg" /><figcaption>Generated by DALL-E</figcaption></figure><p><em>I’m Vinit Tople, an Ex-Amazon Product Leader (in AI &amp; eCommerce) and now an Industry Speaker, Writer &amp; Strategist on AI. Feel free to connect with me on </em><a href="https://www.linkedin.com/in/vinit-tople-9870a43/"><em>LinkedIn</em></a><em>. Read my latest article on AI Differentiation </em><a href="https://medium.com/aimonks/which-organizations-will-win-the-ai-game-differentiation-in-the-age-of-commoditization-c5c71f8d7e7f"><em>here</em></a><em>.</em></p><p>Jeff Bezos was known for keeping an empty chair in meetings to represent the customer. This chair served as a visual reminder of customer obsession, prompting someone to ask, “What would the customer say about this?” It forced employees to think from the customer’s perspective, acting as the customer’s hypothetical proxy.</p><p>With the rise of LLM-powered agents, we are witnessing a significant leap forward in this concept. Companies can now create ‘customer personas’ — internal agents capable of analyzing vast amounts of customer data, from reviews and interactions to business plans and competitive offerings. Over time, these personas will offer real-time insights with a depth and breadth that far surpasses what any human employee could match. Imagine having a permanent fixture (akin to Alexa) which participates in meetings as a customer. The key point? Customers can now speak for themselves, making the empty chair more than just the figurative (perhaps literally?) presence of the customer. It now has a “butt in seat” — a brain and a voice, really.</p><p>But the purpose of this article isn’t just to breathe life into the empty chair. It’s also to discuss an important question many companies face when adopting LLMs: <strong>Which use case should we prioritize in our first wave of investments?</strong> The Customer Persona emerges as a leading candidate, and here’s why:</p><h3>Value Proposition</h3><p>Historically, customer data has been scattered across departments, locked in reports, or accessed only through time-consuming processes. The Customer Persona changes all of that.</p><ul><li><strong>Real-time Insights</strong>: Imagine every strategic conversation, product brainstorm, or marketing campaign enriched with direct, relevant, and real-time customer insights. No need for days of analysis. Customer perspectives can now be injected immediately into decision-making processes.</li><li><strong>Better-Informed Decisions</strong>: For executives, this is more than just speeding up decisions — it’s about making smarter, data-driven decisions.</li><li><strong>Precision &amp; Personalization</strong>: The Customer Persona leverages not only internal company data but also external factors such as competitor offerings and industry trends. Instead of guessing what your customers might want, you can know with precision.</li><li><strong>Scalability</strong>: This persona isn’t static. It can be rapidly updated and scaled as new data sources become available, ensuring you’re always acting on the most current customer sentiment.</li><li><strong>Increased Confidence</strong>: By drawing insights from vast datasets, executives gain confidence in high-stakes decisions. The comprehensive understanding of your customers significantly reduces risk, making it invaluable in navigating today’s competitive landscape.</li></ul><h3><strong>The Permanence</strong></h3><p>Whenever Jeff Bezos was asked, “What’s going to change in the future?”, he would respond with, “I look for what’s <em>never going to change</em> and invest there, instead.” The need for a deep understanding of customers falls squarely into this category. No matter how much AI or market conditions evolve, companies will <strong>always</strong> need to know their customers intimately — their preferences, behaviors, available options, complaints, opinions, and shifting patterns.</p><p>It’s not a question of <strong>if</strong> companies will adopt such personas, but rather a question of <strong>when</strong>, <strong>how</strong>, and <strong>how quickly</strong>. The absence of timely investment in it could in fact pose significant business risk. Given the increasing complexity of customer data and the competitive advantages it unlocks, this is a “no-brainer” investment that aligns with the timeless business principle of customer obsession.</p><h3>Barrier to Entry</h3><p>For many companies exploring LLM applications, the challenge lies in the level of accuracy required and the appetite for risk. LLMs still face well-documented accuracy challenges, including hallucinations. If your use case requires near-perfect precision, LLMs may not be suitable (not yet) without heavy investments. However, the Customer Persona is an excellent fit for LLMs because:</p><ul><li><strong>Internal-facing Tool</strong>: Since the Customer Persona is an internal tool, the risk of backlash from occasional inaccuracies is minimal. It’s for internal decision-making, not direct customer interaction, which lowers the stakes.</li><li><strong>Directional Accuracy</strong>: Even insights that are directionally accurate mark a vast improvement over the status quo. Many companies have customer insights trapped in unstructured data — reviews, transcripts, customer interactions, and internal reports — that are rarely surfaced in real time. Unlocking just a fraction of this data is a huge leap forward, even if it’s not perfectly accurate. .</li></ul><p>This lower accuracy threshold has two key advantages:</p><ul><li><strong>Get Started Immediately</strong>: You don’t need a fully fleshed-out, ideal customer persona to start. Yes, there may be a day with an device with an actual voice replacing the empty chair. But in the meanwhile, with the data you already have — transcripts, reviews, internal reports, and data scraped from forums — you can begin building a useful customer persona today. Even incremental steps toward this vision offer significant value. Perfection shouldn’t stand in the way of progress.</li><li><strong>Plug-and-Play Solutions</strong>: In the past 20 months, RAG (Retrieval-Augmented Generation) approaches have advanced significantly. There are now plug-and-play options that can ingest your existing data and deliver results quickly. These tools offer low-hanging fruit, allowing companies to get started quickly and easily.</li></ul><h3>The Road Ahead</h3><p>The Customer Persona is just the beginning. As AI continues to evolve, we will see this concept extend into other areas of business. As was evident through the Google event last week (Gemini At Work), we’re already seeing the proliferation of agents — customer facing, employee-facing, creative agents, data agents and more. The possibilities are endless.</p><p>Consider the landscape of agents today, much like the mobile app landscape in 2007 when the iPhone launched. Few could have predicted the number of apps we would have today, and the same applies to agents in 2024. As Microsoft AI’s CEO, Mustafa Suleyman said — “we’re building a digital species.” And, these don’t necessarily take 9 months to be born.</p><h3><strong>Key Takeaway for Executives</strong></h3><p>There are use cases with permanent, long-term applicability, as well as quick, scrappy solutions that don’t require high accuracy to deliver value. Many of these use cases are ready for immediate investment, and the Customer Persona is a prime example — a no-brainer. It’s a powerful starting point with low risk and high reward, and it’s only the tip of the iceberg. The opportunities for leveraging AI in your business extend far beyond this, but the Customer Persona is a great place to begin.</p><p>===================================</p><p><strong><em>If you agree/disagree, I’d love to hear your opinion either way</em></strong>. — Vinit Tople</p><p>Also, here are couple other related articles:</p><p><a href="https://medium.com/aimonks/which-organizations-will-win-the-ai-game-differentiation-in-the-age-of-commoditization-c5c71f8d7e7f">Two Crucial Differentiators in AI — While Commoditization Beckons | by Vinit Tople | 𝐀𝐈 𝐦𝐨𝐧𝐤𝐬.𝐢𝐨 | Dec, 2024 | Medium</a></p><p><a href="https://medium.com/aimonks/why-executives-should-make-a-note-of-ai-agents-40dbdc1a2977">Why Executives Should Take Note of AI Agents</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eae47e2ad026" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Two Reasons Why AI is NOT “Just another Invention”]]></title>
            <link>https://medium.com/aimonks/is-ai-just-another-invention-9ca5ee85cd7a?source=rss-3c600bf0ea6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/9ca5ee85cd7a</guid>
            <category><![CDATA[technology-news]]></category>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology-strategy]]></category>
            <category><![CDATA[executives]]></category>
            <dc:creator><![CDATA[Vinit Tople]]></dc:creator>
            <pubDate>Mon, 03 Jun 2024 01:06:15 GMT</pubDate>
            <atom:updated>2024-11-12T19:57:03.774Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*luLeWL3bcKNEWH5frhFflA.png" /></figure><p>Remember when the question “Isn’t AI just another invention?” seemed profound? That was just a year ago. The person who asked that question came off as the voice of reason in a room filled with hype over the latest invention. It was actually a fair and mature question to ask at that time. Today that question sounds innocent at best but not something we’d expect from anyone who’s been paying attention. Let’s dig into why this question has morphed from insightful to quaint and why AI isn’t just another invention.</p><h3>What’s Different With AI?</h3><p>Though AI, like other inventions, will create new jobs and governments will step in to control some of the adverse effects, there are two aspects unique to AI that have never applied to any of the prior inventions and those are the following:</p><p><strong>1) Infinite Inventor, not an Invention</strong></p><p>When the steam engine train was invented, were there worries that the train would not just go off the rail but also build new rails on paths it chooses and enter any area it wishes to? Or that the train increases or decreases its speed on its own? Or that it stops or starts wherever and whenever, as it pleases — without human instruction to do so or any human intervention beyond the initial stage. No, there was no such fear. Neither with the steam engine nor with any other invention whatsoever. That’s where AI is different. AI can think for itself and make decisions for itself. Moreover, it learns concepts or capabilities we (humans) never asked it to learn. Yes — it learns to do things we never taught it to or asked it to. It can recursively innovate itself and build new capabilities. In that sense, as Microsoft AI’s CEO put it, it’s an autonomous and infinite inventor and not just one invention. The key word there is “autonomous”. Autonomy eliminates human intervention, which is great in many cases but absolutely not in many situations. That’s where AI is not just another invention. It has strong potential to go out of human control. We are reliant on processes and good intentions for keeping things under control. Barring nuclear arms, there is no equivalent of processes or human intentions containing proliferation of any technology. And containment of nuclear proliferation had many elements which don’t apply to AI. For example, AI has arguably already proliferated beyond a point of no return, making it far easier to apply than nuclear fusion, which was largely restricted to select governments in the world, when the containment began.</p><p>2) <strong>The Speed of Proliferation</strong></p><p>Trains, phones, computers, and even the internet took decades to pervade through businesses and have a jobs and subsequent economic impact. There was no single train, phone or computer or some internet application which applied to all industries on the day of that invention’s launch. That’s not the case with AI. ChatGPT, since its day of launch, has been ready for application across many tasks across industries, across the globe. Yes, it’s not ready for many tasks and certainly not ready for complete elimination of tasks. But it’s silly to think that only complete elimination of tasks is disruptive. If ChatGPT can create 5% efficiencies across 10 processes in a company or 30% efficiency in one major process, that’s a non-trivial impact, when measured collectively across businesses or industries as a whole. Furthermore, original ChatGPT launch is old news now and a plethora of newer and more potent AI capabilities are chipping away at driving efficiencies (a.k.a job eliminations), at a breakneck speed. That’s a huge impact for a breakthrough that happened just 18 months ago. These “efficiencies” will compound even in just 3–5 years, let alone decades. Yes, eventually new jobs will be created but is society ready to absorb this rate of change? The key word is “rate”. It’s not ready and that’s where AI’s impact will be different. Unlike other technologies, we don’t have a decade or more available for transition. Yes, society today moves much faster than the one when the steam engine came out. Every invention has spread faster than the previous. Even then, the last major job-impacting-general-purpose-technology — the internet — gave the world a decade to adapt. 2–5 years, which AI might give for major impacts, is arguably pushing it too far too soon.</p><h3>Beyond Invention: AI as a Catalyst for Change</h3><p>In the early days, equating AI to just another invention was almost forgivable as we were trying to understand its potential. However, as AI continues to evolve at breakneck speed, it’s clear that this comparison misses the mark by a mile. AI is not merely an invention but more like an infinite inventor, capable of recursively creating new possibilities, technologies, and paradigms — autonomously. Its impact is profound, introducing not just positive advancements but also significant ethical dilemmas. To fully appreciate AI, we must think beyond the traditional framework of inventions. AI represents a leap into a new paradigm where AI models are partners in innovation, not just an invention in themselves. The question, “Isn’t AI just another invention?” needs to be put to rest. We must shift our focus from questioning AI’s relevance to understanding how to responsibly harness its vast potential, acknowledging its transformative force reshaping our world.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9ca5ee85cd7a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aimonks/is-ai-just-another-invention-9ca5ee85cd7a">Two Reasons Why AI is NOT “Just another Invention”</a> was originally published in <a href="https://medium.com/aimonks">𝐀𝐈 𝐦𝐨𝐧𝐤𝐬.𝐢𝐨</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Executives Must Take Note of AI Agents]]></title>
            <link>https://medium.com/aimonks/why-executives-should-make-a-note-of-ai-agents-40dbdc1a2977?source=rss-3c600bf0ea6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/40dbdc1a2977</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[executives]]></category>
            <category><![CDATA[ceo]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Vinit Tople]]></dc:creator>
            <pubDate>Thu, 23 May 2024 06:36:24 GMT</pubDate>
            <atom:updated>2025-01-07T17:31:25.394Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F-CvMGWr-8GHuFpDnfaf8w.jpeg" /></figure><p>As executives evaluate the implications of AI’s rapid evolution, many are cautious. While some acknowledge the potential, they believe the risk-reward ratio isn’t favorable enough to justify significant investments or strategic shifts. In their view, AI remains too unpredictable, and they prefer a “wait and see” approach, investing only in small-scale experimental initiatives. After all, there are still many questions with no clear answers.</p><p>However, there have been critical developments in the field that executives need to consider. These developments have introduced a promising approach to mitigate the risks of AI adoption, making AI more viable for broader applications. Central to this approach is the concept of AI Agents, which are designed to address many of the current limitations in AI systems. By understanding how these agents work, executives may rethink their AI investment strategy. Let’s explore how this solution addresses the key risks.</p><h3>What were the risks and gaps?</h3><p>AI’s potential is vast, but several risks and limitations have hampered its broader adoption, especially for business-critical processes. Some of the most pressing issues include:</p><ul><li><strong>Hallucination</strong>: AI models sometimes provide inaccurate or entirely fabricated answers, not grounded in any authoritative source.</li><li><strong>Bias</strong>: Since large language models (LLMs) are trained on vast amounts of web data, they can inadvertently propagate biases present in that data.</li><li><strong>Mistakes on simple tasks</strong>: Despite their sophistication, AI models often fail to perform basic tasks, like simple math, or may lack common sense.</li><li><strong>Struggles with multi-step tasks</strong>: Many business processes require a series of interdependent steps, something traditional AI models have struggled to manage effectively.</li></ul><h3>The New Solution: AI Agentic Workflow</h3><p>It’s important to recognize that much of the criticism directed at LLMs has focused on their failure to handle basic tasks consistently, such as simple math or retrieving exact facts. Ironically, these shortcomings often overshadow the truly groundbreaking capabilities they offer. In many cases, we’ve been judging AI on tasks for which we already have reliable solutions, rather than leveraging its unique strengths.</p><p>As a sports analogy, consider an athlete who excels at an extraordinary skill, like dribbling in soccer, but struggles with simple tasks like remembering stats or planning his schedule. Rather than dismissing his talent, a good coach would focus on leveraging what the athlete does best and leave the rest to support staff.</p><p>The new approach, centered around Agentic workflow is akin to that support staff. It addresses the limitations of LLMs about getting basic tasks right (among other real challenges) by focusing on where AI excels and supplementing it with tools (built as AI tools or just traditional applications) to handle the rest. Here’s how it works:</p><ol><li><strong>Tools for specific tasks</strong>: AI models may not be reliable for certain sub-tasks, like math calculations or retrieving factual data. Here, specialized APIs (application programming interfaces) act as tools to perform these tasks reliably. For example, instead of relying on an AI model to guess a fact, the model can call upon an API to retrieve verified information from a trusted / authoritative source. For math calculations, there are APIs which can fetch the right calculation. The list goes on but the gist is — don’t use AI for tasks which traditional programming can handle. Instead focus on identifying/isolating such tasks and leveraging the APIs to address that.</li><li><strong>Reflection to reduce bias</strong>: In this approach, AI models are evaluated by another AI model. When the primary model provides an answer, a second model checks it for appropriateness and gives feedback. The process continues until the result meets a predefined standard. This technique, known as “reflection,” reduces bias and increases the reliability of responses.</li><li><strong>Planning for multi-step tasks</strong>: Although this aspect is still in development, it shows promise. The idea is to teach AI models how to break down complex tasks into a series of steps, adjusting the plan as new information becomes available. This capability will be critical for automating business processes, such as determining product pricing based on competitor data, inventory levels, and marketing plans.</li></ol><p>Together, these AI Agents significantly reduce the risks associated with AI. Early expectations for AI were often unrealistic, with the technology being judged for tasks that it wasn’t designed to handle. But now, with AI agents supplementing AI models for specific tasks, the focus shifts toward utilizing AI for areas where it offers truly groundbreaking capabilities, while relying on existing tools for more routine functions. This combined approach is known as the agentic workflow.</p><h3>Conclusion</h3><p>The agent-based approach to AI is not just promising, it’s inevitable. In the near future, AI agents will be as ubiquitous as apps are today. These agents can range from simple tools, like calculators or data retrieval systems, to more complex business applications like pricing assistants. The foundation for this approach is solid, with support from leading AI companies and researchers.</p><p>The most important takeaway for executives is that there is now a clear path to mitigating the risks associated with AI. While not all applications are risk-free, many key business processes can now be automated with much greater confidence. However, building the necessary infrastructure, such as proprietary APIs, will require time and investment. For those who have been cautious or have taken a light-touch approach to AI, now is the time to reassess. The developments in AI Agents open up new possibilities for AI’s role in business, and those who adapt early will be best positioned to take full advantage of its potential.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=40dbdc1a2977" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aimonks/why-executives-should-make-a-note-of-ai-agents-40dbdc1a2977">Why Executives Must Take Note of AI Agents</a> was originally published in <a href="https://medium.com/aimonks">𝐀𝐈 𝐦𝐨𝐧𝐤𝐬.𝐢𝐨</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Exactly Gap Did AI Fill? Machine Learning In Simple Words]]></title>
            <link>https://medium.com/@vinittople/simplest-explanation-of-why-ai-and-what-s-machine-learning-ba0e5cc1fbb3?source=rss-3c600bf0ea6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/ba0e5cc1fbb3</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[machine-learning-ai]]></category>
            <dc:creator><![CDATA[Vinit Tople]]></dc:creator>
            <pubDate>Tue, 14 May 2024 05:07:28 GMT</pubDate>
            <atom:updated>2024-12-23T00:40:11.125Z</atom:updated>
            <content:encoded><![CDATA[<h3>What Exact Gap Did AI Fill? Machine Learning In Simple Words</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*v2hot_O42OCwjWeOO7npgg.png" /></figure><h4>Who is this article for?</h4><p>This article is for someone who wants a non-technical way to understand why/when AI is needed in the first place and how Machine Learning (which is a type of AI) solved for the needs. Machine Learning is one of the foundations of AI. The next level of AI is Deep Learning and even more advanced level of AI is LLM or Large Language Model. Those advanced topics will be explained in separate blogs.</p><h4>Let’s start with the “Why” — Why and When exactly is AI needed?</h4><p>Artificial Intelligence, like any technology, is meant to automate tasks to save time and money and add comforts and conveniences. Broadly speaking, there are two types of tasks — tasks that are well-defined and those that are not. The “well defined” tasks can be automated with a software program but when it comes to tasks that are not “well-defined”, software programming falls short and that’s where we need “Artificial Intelligence. To build that “intelligence”, there are multiple techniques and one of them is Machine Learning. We will get to that technique but first let’s understand what makes a task “well-defined” vs. one that isn’t. To understand, let’s use the example of a bank from the 70s when all the banking processes were still fully manual.</p><h4>Processes which are “well-defined”</h4><p>(<em>Jump to next section if you know how traditional programming can solve for well-defined tasks</em>)</p><p>A well-defined process means one which has a clear set of instructions. There is no involvement of any “guess work” or “estimation, prediction, forecasting” or “any visual inspection” or any “gut check” in the set of instructions. There are clear instructions which are used to process a given set of provided inputs and a clear set of outcomes (outputs). Examples are some tasks that a bank teller used to perform in the 1970s. For example, if a customer walked in, then as “inputs”, they provided the following — the type of request (say cash withdrawal), details like account number/amount, and an ID for verification. The teller was “instructed” to do the following — verify the ID, if verified, then check their bank paper files to see if there was an account with that number and if yes, check if it had enough balance and if yes, then disburse the cash and reduce the bank balance by that amount on some paper form. Similar “instructions” were defined for depositing checks or providing account balances. Barring the “ID verification” which required a visual inspection and hence a “gut check”, each of the other steps were clear with no ambiguity. They could be written down in a “if X, then do Y” type of instructions. Software programming is basically about automating these kinds of tasks. Since the 70s, this entire Information Technology revolution has largely been about automating tasks like these in every industry that’s out there. For example, retail shopping got automated because you just provide an item name, quantity, your address and payment information as input then the software finds a matching price with discounts, adds tax, charges the card and instructs the warehouse to ship the item. Each of those steps is split down into a smaller and smaller but very fixed and clear set of instructions to make all of that happen. Such examples are the “well defined” tasks which can be automated with traditional software programming, without the help of any AI.</p><h4>Processes which are NOT “well-defined”</h4><p>Banks do a lot more than these “well defined” tasks. For example, bankers approve loans for homes. Bankers also provide investment counseling — where to invest, what the risk is and so on. Bankers also process an ad hoc set of requests through their “customer service” counter. Each of these tasks has several elements of “gut check or judgment or estimation or prediction” involved — the aspects that traditional software programs cannot handle.</p><p>Let’s start with the first scenario of Loan Application. Let&#39;s say a customer requested a loan and provided all his information like salary, household income, net worth, number of dependents, other open loans etc. Now the loan officer has to look at all these factors and estimate the probability of this customer defaulting on the payment. If the probability is not extremely low (it can never be 0%), then loan will be approved. Back in the 70s, the loan officer had to “guess or estimate” that probability based on his gut or judgment. Now in the 80s/90s, after automating the well-defined tasks of the Bank Teller, the software engineers asked the Loan Officer — “Can you describe steps you follow so I can write a program to automate those steps for you”. The Loan Officer replied saying — “<em>I look at all the information in the Loan application and then use my judgment to estimate how likely the customer is to default</em>”. The software engineer said — “<em>Well sir, there’s no way to put your judgment in a software program</em>”. The result was that the Loan Application processing remained completely manual, unlike the teller tasks like “cash withdrawal” or “account balance inquiries”.</p><p>Same thing happened when the software companies tried to automate the job of the investment counselor in the bank. The investment counselor said “<em>I look at stock market, the company financials and some charts and then the risk profile of the customer, their retirement age and their investment goals and then use my judgment to recommend an investment plan</em>”. The software engineer asked — “<em>why don’t you follow a set of rules to recommend investments</em>” and the Counselor said “<em>I can’t because there’s way too many factors. Factors about the customer, the companies to invest in, where the stock market level is, the current economy, the currency exchange rate, the political news and so much more. How do I make a rule around all of that? Even if I do, it&#39;s still all a guess. My best guess</em>”.</p><p>The situation was even worse with the customer service (CS) desk. When the software engineers told the CS guy to describe the rules or instructions he follows to do his job, he said, “<em>I just use common sense. People come here with all sort of things. One man came here panicked because of a charge on his account statement. Turned out it was by his son. One woman came and talked about her wedding for a while and I had to just guess that she wanted her maiden name changed on the account. Another person wanted to block his estranged wife from having access to their joint account. These are all so random scenarios. How can I give you a clear set of instructions. You just have to use common sense.</em>” Clearly, this Customer Service process couldn’t be automated either.</p><p>In a nutshell, traditional programming falls short when there are no clear set of instructions or when there is prediction or estimation or judgment required. This typically happens when there are way too many factors involved in a decision and when there are very complicated relationships between them which can be put down in a simple “if X, then do Y” type of steps.</p><h4>That’s Where AI Steps In</h4><p>The Software Engineers left the bank saying — “<em>we automated all your well-defined tasks but for the rest — you guys keep your people because we can’t replicate human judgment in a program”</em>. This is when the AI guys stepped in to say — “lets give this a second shot”. AI guys means Data Scientists. They reached out to the Loan Officer and said, “<em>Hey, you say that this is all your judgment, but can you explain how you built that judgment? When you started your career, you didn’t have this judgment, right</em>”.</p><p>The Loan Officer said — “<em>Of course not, I didn’t know anything at that time. It&#39;s been 35 years for me at this bank and I have built this judgment over these 35 years. As I started approving loan applications, I would follow their mortgage payments on the subsequent years and especially started looking at all the people who defaulted and couldn’t make the loan payments. I started looking at what was common between all these defaulters and observed a few patterns. For example, a lot of defaulters were the ones who had defaulted on their credit card payments or other bills. But when it came to marital status, singles were likely to default more but just a little bit more than married couples. This told me that some factors carry more importance than others in predicting the default possibility. What I mean is that the marital status didn’t matter much but bill payment history said a lot. This way I’d keep refining my thinking about a variety of factors and in my head, I’d form a judgment about what was important and what wasn’t. That’s it — I’d then have these judgments at the back of my mind while processing the applications and I’m known for being fairly accurate in my predictions.”</em></p><h4>This is where the Data Scientists said —</h4><p>“<em>Well, this sounds more like math than some magic which only humans can do. We can make a computer do this!</em>” And what they came up with as a solution is now called “Machine Learning”. So, what is it exactly?</p><p>To the Loan officer, the Data Scientists said — “<em>Give me a list of all the customers who took a loan from the bank along with all the details collected during their loan application (like their salary, their history of missing payments on bills, their salary, the list of all their loans, their family details etc.) and also tell me which of these customers landed up defaulting on the loan</em>”</p><p>When they got this list, they wrote a computer program to calculate how frequently a ’history of missing payments on bills’ was already on the list of defaulters at the time of their loan application. That turned out to be the case in 80% of the defaulters. Clearly, if this factor had been given enough weight, a lot of defaults could have been avoided. Then the program found that “single” marital status showed up in only 20% of the defaulters, which indicated that marital status was not really a helpful indicator. They did this for all possible factors collected on the application and now it was clear which factors mattered more and which mattered less. In other words, the computer program now had the same insight that the Loan officer was using to make his decisions. To ‘store’ this learning, the Data Scientists asked the computer to assign different levels of importance to each factor based on frequency at which it appeared on the list of defaulters. This level of importance was called the ‘weight’ of that factor and it was basically a number between 1 to 10 where 10 is a very strong indicator of default and 1 is very low indicator of default. For example, when a person had a history a late or missed payments, the computer would add a score of 8 (out of 10) to the customer’s ‘default probability score” and if someone was single, then it would add just 2 to their score. When all such ‘weights’ were added up, the loan applicant would get that final default probability score. The higher the number, the higher the chances of default. This computer program which memorized all these ‘weights’ against each of the factors (also called attributes) was called the ‘Machine Learning Model’. And because Machine Learning is a basic form of AI, its sometimes also called the ‘AI Model’.</p><p>So, the eventual solution for the loan application process looked something like this: When a customer filed a loan application, all the details (like salary, history of missed payments etc.) were the ‘inputs’ to the ML model. The ML model would apply the right ‘weight’ to each of the input factors which would generate the “default probability score”. One aspect worth noticing about this ML based solution is the presence of this additional ‘weight layer’ between the input and the output layers. In other words, presence of just two layers (input &amp; output) is what traditional programming is about while ML adds that third layer of weights.</p><p>Anyway, so in our example, the Data Scientists went to the Branch Manager to say that their ML model can now replicate the judgment of the Loan Officer. In fact, the ML Model is more accurate because it can remember and calculate these complicated details more accurately. In this simple example, this approach of “Machine Learning” was predicting the probability of loan default with the same level of accuracy as the Loan Officer and perhaps even better than some other Loan Officer who are less diligent than others.</p><h4>Summary and Conclusion</h4><p>In summary, traditional programming can solve tasks which are well defined. You give a set of input, and the computer will give you output based on a set of ‘rules or instructions’ which look like “If X, then Do Y”.</p><p>When the task is not well defined or when you need judgment or ability to predict or estimate — traditional programming cannot handle it. To be able to do handle those scenarios, a computer first needs to analyze data and learn how much each of the input matters and to what extent. This ‘extent’ is stored as a ‘weight’ of that particular factor and the collection of such weights is the ML model. The ML model applies these weights to each of the inputs to make the prediction or the estimate. This approach of learning from data to predict or estimate is called as Machine Learning.</p><p>Lastly, while this ML approach can handle some business scenarios like the Loan application scenario, it falls short of giving adequate level of accuracy to handle the ‘investment counseling’ process of the bank mentioned earlier and is even less accurate when applied to the customer service process in the bank. For those scenarios, we need more advanced form of learning to build the right AI model and that technique is called Deep Learning. A similar article on Deep Learning to follow soon.</p><p>If you found this article helpful or interesting, please Clap or Follow!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ba0e5cc1fbb3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Other LLM in AI — Leaders, Lawyers and Marketers]]></title>
            <link>https://medium.com/ai-insights-for-executives/the-crucial-role-of-leaders-lawyers-marketers-in-gen-ais-gtm-strategy-0a244336850b?source=rss-3c600bf0ea6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/0a244336850b</guid>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Vinit Tople]]></dc:creator>
            <pubDate>Thu, 09 May 2024 01:40:39 GMT</pubDate>
            <atom:updated>2024-09-17T05:39:11.463Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Crucial Role of Leaders, Lawyers and Marketers in the Age of AI</h3><p>The Consumer Electronics Show 2024 in Las Vegas was abuzz with one overarching theme: AI. It marked the culmination of a year filled with AI powered prototypes across industries, showcasing the remarkable advancements in AI through Large Language Models (LLMs). The spotlight quickly shifted to the next step — the path from the prototypes to real products. While the prototypes demonstrated the magical power of LLMs, this final step to the real product is ridden with landmines — the risks and the harms of LLMs with all their inherent unpredictability. This article aims to 1) shed light on these challenges, 2) outline common approaches to address them, and 3) propose recommendations with supporting rationale. But let’s start with some background first.</p><h3>Background</h3><p>Traditional software programming faced limitations in tasks involving natural language processing (NLP). Examples included customer support via chatbot, product feedback analysis, voice- operated smart home, language translation for global websites, among others. Although traditional AI addressed these, it struggled with complex conversations, understanding only short and simple sentences. (e.g. how do I cancel my order or turn on the porch light). The advent of LLMs in 2022–23 bridged this gap, almost achieving the elusive goal of true natural language understanding. Suddenly, longer sentences and even sentiments were being understood by AI, driving widespread excitement. But then, the limitations came to the fore soon. The LLM behavior was inconsistent. While it would understand sentences you didn’t expect it to, it would fail to understand sentences you expected it to. The unpredictability was jarring. Secondly, LLMs could even generate content (not just understand), but the generative ability is a double-edged sword. When it’s leveraged for writing poems or songs, it is hailed as creativity but when answering a legal/medical/business question, the same creativity came to be known as hallucination (making stuff up). That’s not all because there were other issues. The whole of the internet was used to train the model so there were concerns of copyright violation. Interactions with LLMs were being used to train/refine them further so that raised privacy concerns. In short — the list of yellow and red flags was long.</p><h3>The Catch-22</h3><p>All the challenges aside, there is no denial that tremendous value lies in the LLMs. There is something on offer for every company and for a multitude of use cases. The industry is brimming with ideas to unlock it — from productivity gains to process automation to analytics to novel customer facing features to innovative business ideas. The key challenge here is that the significant value of LLMs comes as a package deal with a non-trivial bundle of risk. That leads to the catch-22, which most leaders and companies are grappling with — not adopting the technology presents a serious opportunity cost and adoption is akin to walking into an area with landmines.</p><h3>Industry response</h3><p>Industry responses to this dilemma vary based on a variety of factors. While some companies opted to observe from the sidelines, awaiting technological maturity, others decided to actively engage with LLMs. Among the latter, two predominant approaches emerged and that is the focus of this article.</p><p>The first approach centers on establishing core LLM teams comprising of data science, engineering, and product functions only. The predominant focus of this approach is to reduce the risks and limitations of LLMs. Efforts include injecting ethics into LLMs, reducing bias, mitigating hallucination, validating knowledge, tightening privacy controls, customizing LLMs, and integrating with IoT systems. All of that is a valuable investment, even essential for the most part. However, what’s conspicuous by their absence (or at least the absence of active participation) is the role of functions beyond the core team. Beyond these core functions of science, engineering and product, the rest (legal, sales, marketing, operations etc.) are largely playing passive roles, ready to support as and when the core team is ready to engage them. By ‘ready’, the expectation being that the risks and limitations are brought down to acceptable levels. At the surface, this approach of ‘fix-the-issues-first-before-involving-others’ sounds reasonable, but is it? Some companies are looking at it differently and that’s the second approach.</p><p>The second group has realized that waiting for the core team to bring the risks/limitations down to traditionally acceptable levels is not a pragmatic approach. The resolution or even mitigation of these issues down to the traditionally acceptable level could take a very long time. That’s just how complex these issues are. These are issues such as removing bias, injecting ethics, hallucination mitigation etc. They have also realized that the LLM builders (e.g. Big Tech and open-source communities) are better equipped to make progress on these exceedingly difficult challenges than their in-house teams, which are much smaller. Also, any improvements at the source (Big Tech/open-source community) will trickle downstream to them anyway. In other words, attempts to solve these foundational problems of LLMs with in-house resources has the risk of being a redundant effort or a ‘throw-away’ piece of investment. Recognizing that, these companies have chosen to focus their attention on aspects which <em>only</em> <em>they</em> can and need to solve. This includes steps specific to the company’s use cases like custom prompts, training with proprietary data, data validation against relevant sources, API integrations, relevant context injection, among others. While the focus on only the company-specific issues is one difference from the first approach, the second and the more notable difference is the presence of workstreams beyond science/engineering. That includes the efforts to craft a Go-To-Market (GTM) strategy, specific to their use case(s) in parallel to all the science and engineering efforts. In other words, these companies have adopted a multi-pronged or parallel approach (vs. serial) <em>from the very outset</em>, engaging functions beyond just the core ones. These companies have realized that ‘traditionally acceptable’ state is not a very practical proposition, at least in the short term. Instead, non-traditional paths need to be found beyond just science and engineering; it will take a village to extract value safely from this precocious but perilous technology. More specifically, their leaders, in partnership with the legal and marketing functions, are working alongside the core team to carve a way through these imperfections and even risks.</p><h3>The Right Approach</h3><p>Like most decisions, there is no one-size fits all approach. However, there is good reason to hypothesize that the biggest success stories in the LLM adoption space will come from companies that recognize the difference between the two approaches above and choose the second — the multi-pronged approach from the outset. In fact, most companies in the first group will soon realize that while they kept applying traditional benchmarks of acceptance, a lot of competitors garnered large gains in the meanwhile by accepting the imperfections and working through them. It’s all just a matter of when this realization arrives, and its timing can be the differentiator in this dizzyingly fast race.</p><h3>More on the multi-pronged approach</h3><p>AI will penetrate all the functions of any organization and the multi-pronged approach helps in building the muscle needed to adapt to this massive and imminent change. LLM powered solutions should not be left to the few scientists and techies. Every function will need to adapt and reinvent their roles. Taking the multi-pronged approach for determining the GTM strategy for the prototypes is the first step in that direction and could provide a competitive edge in the medium to long term. More specifically for now, Leaders, Lawyers and Marketers have their task cut out. While the specifics of what these roles need to change or do will vary by situation, below are some thought starters on how these roles will have an elevated significance in the AI era.</p><p><strong>Marketing </strong>needs to recognize that one aspect of their role has an outsized importance in the world of LLMs — it’s to position the product right,<strong> </strong>in terms of setting realistic expectations or even lowering certain expectations. Resisting the urge to sugar-coat limitations is imperative. Instead, transparency regarding product limitations can facilitate adoption rather than hinder it. That’s because the opinions about LLMs have oscillated too dramatically and too quickly, from fervent excitement to deep concern — often leading to excessive caution. A transparent and honest positioning will be a refreshing perspective and alleviate undue skepticism and apprehension, where appropriate. It can also be the differentiator between an acceptable level of backlash vs. outright rejection or even elevated legal risks. ChatGPT is a prime example of getting this just right.<strong> </strong>The<strong> </strong>product positioning (though not explicitly stated) was — “<em>if you want the latest, most accurate information and total privacy, wait for others. In the meanwhile, for everything else, you have us. But use at your own peril.</em>” This product positioning was an integral enabler of ChatGPT’s path to market despite the imperfections and risks. The rest, as they say, is history.</p><p><strong>Legal</strong> teams are not mere support functions in the era of Gen AI. Like marketing, they will play a key role in any LLM powered product’s GTM strategy because the legal implications are one of the biggest holdups on that front. Many LLM powered products linger in prototype stage unresolved legal issues — who owns liability when the product misfires, where are the lines of accountability, privacy, and data protection concerns, could LLM responses hurt your brand, copyright infringement and so on. Many of these are legitimate concerns unique to LLMs. However, there also are questions, unnecessarily jumbled up with LLMs although they’re just a variation of existing concerns. For example, there is fear that giving data to LLM builders (e.g. Big Tech) for model training could result in the data being available to your competitors. That’s not an LLM specific problem — this concern exists even when companies move their IT infrastructure to the cloud or share data with partners. There are answers for this — contractual and technical. Jumbling these issues together is not helpful. Differentiating between these issues is crucial to focus on challenges truly unique to LLMs, a task in which legal functions play a pivotal role. However, to do this effectively, legal professionals must educate themselves on AI and LLMs extensively, stay updated on developments, rethink their approaches, and find the paths forward. Even then, their task — to craft T&amp;Cs for products powered by an unpredictable technology — will remain extremely challenging. However, without this legal contribution, the value in the LLMs and the subsequent opportunity from it will remain trapped. Consequently, though an incredibly challenging task, it’s also a once-in-a-lifetime opportunity to play a differentiating role in this revolutionary technology.</p><p><strong>Leaders</strong> have already been navigating the complexities of the AI revolution, tackling the macro questions for their organizations — what’s in it for us, how do we prepare/contribute and so on. While many leaders have authorized and supported the necessary investments, the savvy ones are one step ahead. Along with the resource investments, they are spearheading cultural changes in their organizations that this new world requires. First cultural change needed is to embrace and normalize ‘trial and error’ as integral to product strategy development, a practice largely limited to only R&amp;D organizations until now. It is a departure from traditional practices where product specs were largely a function of what business and product teams laid out while engineering built it. With LLMs, there is a new brain in the room and what it can and wants to do can be discovered mostly through trial and error, at least as of now. This process is a much larger determinant of the final product than traditional processes. Second, leaders must establish a robust framework for risk assessment across all organizational levels. While traditionally, risk assessment relied on subjective judgment from leaders and lawyers, the complexity of risks associated with LLMs demands a more decentralized approach due to the sheer volume of it. That said, leaving it to a broad swathe of those on the assembly line (engineers, scientists and product managers) will also lead to issues. To ensure consistency in risk evaluation, there is a need for top-down guidance and support. Lastly, when these risks become reality and some setbacks occur, leaders must demonstrate unwavering support and solidarity. Their teams need that air cover to function effectively on this shaky ground.</p><h3>Conclusion</h3><p>The flurry of LLM powered prototypes in 2023–24 demonstrated the enormous power of LLMs. But the vast chasm between the number prototypes and actual products revealed the gravity of challenges and risks in that final step — prototype to real product. While investment in the resolution or mitigation of these risks is essential, that itself is not enough. Companies need to acknowledge that a significantly elevated level of uncertainty, risk and ‘trial and error’ are the new normal and here to stay. Consequently, the work to function in the new normal and get products out in the market even in the presence of some risks/imperfections must begin earnestly too. That will call for innovation not just in technology but also in key functions beyond the technical ones. Leaders, Lawyers, and Marketers (coincidentally another ‘LLM’) are that next line of attack and have their task cut out to define the GTM strategy. The answers they need to come up with aren’t clear yet and may not be for a while. However, these functions will need to begin organizing themselves to be effective in this new paradigm and this will take time. The individuals, leaders and companies that invest in this aspect upfront will build the necessary muscle and reap the rewards sooner than others.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0a244336850b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-insights-for-executives/the-crucial-role-of-leaders-lawyers-marketers-in-gen-ais-gtm-strategy-0a244336850b">The Other LLM in AI — Leaders, Lawyers and Marketers</a> was originally published in <a href="https://medium.com/ai-insights-for-executives">AI Insights for Executives</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>